Data centres and cloud

What it is about

A data centre is a building whose main job is to deliver electricity and cooling to computers and to connect them by network. In an AI data centre, thousands of GPUs are wired together so tightly that they work on one task like a single machine. According to the International Energy Agency (IEA), the servers take about 60% of a data centre's electricity on average; the rest goes into cooling, storage, network and backup power.

Few companies own such a building. Most rent compute through cloud computing, from hyperscalers such as Amazon, Microsoft and Google, from specialised neoclouds or as a model behind an API. With that choice you also decide what you pay, where your data is processed, whose law applies to it and how easily you can switch later.

What a data centre physically is

Picture a large warehouse with a power plant and a refrigeration system attached. Electricity arrives from the grid through transformers, passes through uninterruptible power supplies (UPS, batteries that bridge short outages) and ends in the racks: steel cabinets that hold servers, storage and network switches. Backup generators stand by for longer outages. Almost every watt the servers draw becomes heat, and the cooling system has to carry it away. Software for scheduling, monitoring and maintenance turns this hardware into capacity that customers can use.

The IEA gives the typical split of electricity inside a data centre. The share for cooling shows how efficient a site is, and it is the main driver of PUE, the ratio of total site electricity to the electricity that reaches the IT equipment.

Reliability is graded by redundancy. The Uptime Institute's Tier Standard has four levels. Tier I has basic backup and must shut down for maintenance. Tier II adds spare components such as generators, UPS modules and chillers. Tier III can maintain any part without switching off the servers. Tier IV runs independent, physically separate systems, so that a single event cannot take both down. The official standard defines these levels by design and names no uptime percentages; Uptime also sells the certification.

• Servers: about 60% of the electricity on average, with CPUs, GPUs and memory. The share varies by type of data centre (IEA).
• Cooling: about 7% in efficient hyperscale facilities, more than 30% in less efficient enterprise data centres (IEA).
• Storage and network: about 5% for storage and up to 5% for networking (IEA).
• Backup and building: UPS batteries and generators are rarely used and still required for reliability. Lighting and offices add a small share (IEA).

Storage and software at AI scale: for training Llama 3, Meta used a distributed storage system with 240 PB on 7,500 servers with SSDs. To keep 24,576 GPUs busy, it had to rewrite job scheduling, network routing and communication software; before these fixes the utilisation of large clusters ranged from about 10% to 90%, afterwards above 90% (Meta, vendor report).

How an AI data centre differs from a classic one

The difference starts in the rack. The most frequently reported rack density in the Uptime Institute's 2026 survey is 11 kW, up from 9 kW in 2025. NVIDIA's GB300 NVL72, a rack with 72 GPUs, draws up to about 142 kW according to NVIDIA's reference architecture, as reported by Data Center Knowledge. Air cooling becomes impractical above roughly 50 kW per rack, so AI racks use liquid cooling: cold plates sit directly on the chips (direct-to-chip) and handle 100 to 150 kW per rack. According to the vendor Schneider Electric, this method held about 55% of liquid cooling in 2026. The power supply changes too. The classic 54-volt distribution inside a rack tops out at about 200 kW, so the industry is moving to 800 volts direct current.

The second difference is the network. Inside the rack, GPUs talk over a very fast private link: in NVIDIA's GB200 NVL72, 72 GPUs and 36 CPUs form one NVLink domain with 1.8 TB/s per GPU and 13.4 TB of HBM memory in total (vendor data). Between racks, an interconnect links thousands of such units into a cluster. Two technologies compete: NVIDIA's InfiniBand and Ethernet with RoCE. In 2024 Meta built one cluster of 24,576 GPUs with each and trained Llama 3 on the Ethernet cluster, by its own account without network bottlenecks.

The third difference is failure. During a 54-day window of training Llama 3 405B on up to 16,384 H100 GPUs, Meta counted 466 interruptions, 419 of them unexpected. About 78% of the unexpected ones were hardware problems, most often faulty GPUs (30.1%) and HBM memory (17.2%). Effective training time still stayed above 90%, because the job saved its progress regularly (checkpoints) and restarted quickly. Meta also reports that thousands of GPUs changing their power draw at the same moment made the site's consumption swing by tens of megawatts, a strain on the grid (vendor report).

Free floor space alone makes no AI site. Power per rack and cooling set the limit, which is why many older halls cannot host GPU racks. Anyone renting a large cluster should ask how failures are handled, how much spare capacity exists and which credits the contract grants for outages.

Training clusters, inference clusters and latency

Training a large model runs one job for weeks on thousands of tightly connected GPUs. For Llama 3 405B, Meta used up to 16,000 H100 GPUs in a cluster of 24,000, with 400 Gbps of network per GPU and 240 PB of storage, and thousands of GPUs had to save and reload checkpoints in sync. Network and reliability dominate here. Because nobody waits for an individual answer, a training site can stand far from users, where power and land are cheap.

Inference serves many short requests, and every user waits. Two measures describe the wait: time to first token (TTFT), how long until the answer starts, and time per output token (TPOT), how fast it then streams. Reading the prompt (prefill) and writing the answer (decode) load the chips differently. The research system DistServe, presented at the OSDI 2024 systems conference, runs both phases on separate GPUs and served up to 7.4 times more requests within the latency targets. Vendors define their own targets: NVIDIA's benchmark for the GB200 assumes 5 seconds to the first token and 50 ms per token (vendor data).

Distance between user and data centre adds surprisingly little. Google Cloud's region guide gives about 1 ms of round-trip time per 100 km, with real network paths 1.5 to 2 times the map distance, and 100 to 500 ms on a 3G mobile connection. Applications that tolerate 100 ms or more can usually run in one region. A model that writes for several seconds dwarfs that delay. Our reading of the sources: region choice for AI is mostly driven by law, price, available GPUs and models and the carbon intensity of the grid. Voice and real-time applications are the exception, where every part of the latency counts.

Types of data centres and AI infrastructure

Only some of these terms have an official definition. The US standards institute NIST defines deployment models (private, community, public and hybrid cloud) and edge computing. The EU defines its gigafactory programme and proposes sovereignty levels. Hyperscale, neocloud, AI factory and modular are industry terms with no standard definition or threshold. The working definitions below are an Aulory synthesis of the sources, with figures where they exist.

The balance is shifting. According to Synergy Research, hyperscale operators held 48% of global data centre capacity at the end of 2025, colocation providers 20% and companies' own sites 32%. By 2031 Synergy expects the hyperscale share to reach 67% and own sites to fall to 19% (forecast).

• Hyperscale: very large campuses that a hyperscaler runs for its own cloud and services. No official threshold; Synergy counted 1,360 such data centres at the end of 2025.
• Colocation: in colocation, a landlord rents out space, power, cooling and connectivity, and tenants bring their own servers. Rents are quoted per kW and month.
• Enterprise: a company's own server room or building, the classic form of self-hosting.
• Edge: small sites close to devices and users, for fast answers and local data. NIST describes edge and fog computing in a 2018 model.
• Modular: prefabricated units in containers or on frames, set up quickly. We found no source with a definition.
• Neocloud: a cloud specialised in renting GPUs, often on multi-year contracts. No standard definition; Synergy names CoreWeave, Nebius, Crusoe and Nscale among the fastest-growing providers.
• Sovereign cloud: a sovereign cloud aims to keep location, operation and applicable law in one jurisdiction. No legal definition yet; the EU proposes four levels.
• AI factory and gigafactory: AI factory is NVIDIA's marketing term for a GPU-dense data centre. The EU's AI gigafactories are an official programme of EuroHPC: the call of 30 July 2026 seeks up to seven sites, with applications due on 12 November 2026.
• National supercomputer: state-funded research systems. In the TOP500 list of June 2026, LineShine in Shenzhen leads with 2,198.4 PFlop/s, ahead of El Capitan in the US; JUPITER in Jülich is the first European exascale system with 1,000 PFlop/s.

The TOP500 ranks computers with a benchmark in high double precision (HPL). AI works with much lower precision, so the list says little about AI capacity. Commercial AI clusters such as those in Epoch AI's database appear in it rarely or not at all.

How you rent compute and what you pay

The US standards institute NIST defined cloud computing in 2011 by five properties: self-service on demand, access over the network, shared pools of resources, rapid scaling and metered use. It names three service models. With IaaS you rent raw machines, storage and networks and run everything above them. With PaaS you bring your code and the provider runs the platform. With SaaS you use finished software, such as a chat assistant.

AI adds layers in between. Renting GPUs by the hour, often called GPU as a service, is a form of IaaS. Calling a model through an API and paying per token is often called model as a service; neither term has an official definition. Microsoft's AI platform shows the range: standard deployments billed per token or as reserved throughput, serverless endpoints billed per token, and managed compute, a dedicated endpoint for your own or open models billed per minute of uptime. Labels can mislead. Amazon's SageMaker Serverless Inference scales to zero when idle and bills per millisecond, and it supports no GPUs.

You pay in one of two ways: for time on hardware or for output. Every discount buys away some flexibility. The main forms, with list prices from vendors and providers:

• On demand: pay per GPU-hour, no commitment. In August 2026 an H100 cost US$3.85 (Nebius) to US$6.16 (CoreWeave) per hour at neoclouds, a B200 US$6.69 (Lambda) to US$8.60 (CoreWeave), according to list prices compiled by MarkTechPost.
• Reserved: a commitment for months or years lowers the rate. On Amazon Bedrock, reserved throughput for Cohere Command costs US$49.50 an hour without commitment and US$23.77 with six months, about 52% less (vendor list price).
• Spot: spare capacity at up to 90% off the on-demand price at AWS, which the provider can reclaim with two minutes' warning. CoreWeave listed an H100 at US$2.46 as spot.
• Per token: input and output tokens are priced separately, output usually higher. On Amazon Bedrock, DeepSeek v3.2 costs US$0.62 per million input tokens and US$1.85 per million output tokens in US regions (25 October 2026).
• Batch: requests that can wait are processed later at a discount. Bedrock's batch and flex tiers cost 50% less; the priority tier costs 75% more.
• Dedicated: your own endpoint or cluster, billed for the time it runs. This suits steady load, your own models and strict isolation.

GPU price lists disagree. An earlier compilation by an advisory firm put the B200 at US$4.99 to 6.04 per hour on demand at specialist clouds, below the list prices above. Dates, contract types and regions differ, and negotiated prices are rarely published. Treat every figure as a snapshot.

How big the build-out is

Nobody measures global data centre capacity the same way. Some count the power the IT equipment can draw, others the power a site demands from the grid, others only AI chips. The estimates below measure different things. Never average them.

Among single sites, Epoch AI's list of 20 October 2026 is led by Colossus 2 in Memphis with about 1.1 million H100 equivalents (H100e, a unit that converts different chips into the compute of one NVIDIA H100). It is followed by Anthropic and Amazon in New Carlisle, Indiana (818,000), Microsoft Fairwater in Atlanta (769,000) and Google in Pryor, Oklahoma (637,000). Places five and eight go to two DayOne sites in Johor, Malaysia, the only non-US sites among the top eight. Plans run far ahead of reality: of more than 9 GW planned for OpenAI's Stargate in the US, about 0.3 GW was running in April 2026.

The money comes from a handful of companies. In their own earnings releases, Amazon reported purchases of property and equipment of US$54.2 bn in the second quarter of 2026 and US$173.0 bn over twelve months (gross). Microsoft reported US$115.9 bn for its fiscal year to June 2026. Meta reported US$31.08 bn in the quarter and expects US$130 to 145 bn for 2026. For Alphabet (US$44.9 bn in the quarter, guidance of US$195 to 205 bn for 2026) and Oracle (US$28.45 bn in the quarter to August 2026, up to US$95 bn for the fiscal year to May 2027), we found the figures only in press reports. Companies define capital spending differently, gross or net and with or without finance leases. The IEA puts the capital spending of five large tech firms above US$400 bn in 2025, rising by another 75% in 2026.

Neoclouds grow faster on paper than in concrete. CoreWeave reported 3.7 GW of contracted power at the end of June 2026 and 1.5 GW active. Crusoe states 4.9 GW contracted and about 0.2 GW active; Nebius aims for 5 GW contracted and about 1 GW connected by the end of 2026 (company figures, compiled by MarkTechPost). A lot of this is financed with debt: CoreWeave raised about US$18 bn through loans, convertible bonds and shares in one quarter, and JLL counts US$17 bn of data centre debt sold as securities in the first half of 2026.

The market is tight. JLL puts vacancy in North America at about 1% for the third year running, with more than 66 GW under construction, 77% of it in new markets such as West Texas, Ohio and Louisiana. CBRE counts 7.5 GW under construction in the eight primary colocation markets, more than 80% of it let in advance; the two figures differ in scope by an order of magnitude. In Europe, vacancy in the core markets Frankfurt, London, Amsterdam, Paris and Dublin (FLAP-D) was 7.3% in the first quarter of 2026 (CBRE). In Asia-Pacific, colocation vacancy stood at 10.3% with a record pipeline of 26.5 GW (Cushman & Wakefield).

• JLL (broker): about 103 GW of global capacity at the start of 2026, nearly 200 GW by 2030, with AI about half of it by then (forecast, reported by IREI).
• TrendForce (market research): power demand capacity of 122.9 GW in 2025 and 161 GW in 2026, with AI servers at about a third in 2026; 490.7 GW forecast for 2030.
• Epoch AI (chip-based): about 30 GW of AI data centre power worldwide in the fourth quarter of 2025, estimated from chip numbers times their power, plus overhead.
• Epoch AI (largest sites): the 95 largest tracked AI data centres hold 14.8 million H100 equivalents with 14.2 GW of IT power (20 October 2026).

Where your data sits, whose law applies and how you leave

Two questions are often mixed up. Data residency asks where data is stored and processed. Data sovereignty asks whose law and whose control can reach it. Choosing an EU region answers the first question. The answer to the second depends on who owns the provider, who operates the systems, who holds the encryption keys and how courts read the law.

The US view: the Department of Justice explains that the CLOUD Act of 2018 clarified that providers under US jurisdiction must hand over data in response to valid US legal process, wherever they store it. In its reading the law creates no new type of warrant and allows no bulk collection, stored content requires a warrant approved by an independent judge, and conflicts with foreign law are weighed case by case. This is the US government's own interpretation. Agreements under the act exist with the UK (2019) and Australia (2021); talks with the EU resumed in 2023 without a concluded agreement.

The EU view: in the Schrems II ruling of 16 July 2020, the Court of Justice struck down the Privacy Shield, because US surveillance was not limited to what is strictly necessary and people in the EU lacked effective remedies in court. Standard contractual clauses stayed valid, with a check in each case. Since July 2023 the EU-US Data Privacy Framework again allows transfers to certified US companies; a challenge was dismissed in September 2025 and is on appeal, and we could not verify the outcome. For non-personal data, the EU Data Act requires cloud providers to take technical, legal and organisational measures against unlawful access by foreign governments. Under the GDPR, every transfer of personal data needs a legal basis.

Practice holds traps. Microsoft's documentation states that stored data stays in the chosen geography, while in a global AI deployment prompts and answers may be processed wherever the model runs. Sovereign offers range from European providers to separate partitions of US providers, such as the AWS European Sovereign Cloud in Brandenburg, open since January 2026; its sovereignty is AWS's own claim. The EU's proposed Cloud and AI Development Act of June 2026 would grade sovereign clouds in four levels for public procurement, from hosting in the EU to protection from foreign law.

Switching is getting easier by law. The EU Data Act removes all switching charges, including egress fees, from 12 January 2027; until then providers may only charge costs directly linked to the switch. PaaS and SaaS providers must offer open interfaces and exports in common machine-readable formats, and IaaS providers must help reach comparable results with the new provider. AWS already waives egress on request for customers who leave. Our reading: today's lock-in sits mostly in proprietary services, prompts tuned to one model, fine-tuned models on closed platforms and committed spend.

Switching periods under the Data Act, as summarised from draft guidance by the Dutch regulator ACM (secondary report, October 2026): notice of up to two months, a transition of up to 30 days and at least 30 days to retrieve your data. During parallel use, providers may still charge actual data transfer costs.

Efficiency and sustainability

The standard measure is PUE, defined in the international standard ISO/IEC 30134-2; its 2016 edition was replaced by a new one in 2026. The Uptime Institute's 2026 survey finds an average of 1.52 per site and 1.36 weighted by capacity, because large new sites routinely reach 1.3 or lower. PUE measures the overhead around the chips. It says nothing about the carbon in the electricity, the water used or how well the chips are used. Its counterpart for water is WUE.

Rules are tightening. The EU plans a rating scheme for data centres from 500 kW of IT power, with classes from A (PUE of 1.15 or less) to G (above 1.9) and similar classes for water, and labels generated automatically from 15 August 2027. Sources disagree on the status: the law firm Linklaters describes the Commission act of 21 September 2026 as adopted, the law firm Beveridge & Diamond as proposed amendments under a two-month scrutiny period. Both agree that it is not yet in force. China's industry ministry targets a PUE below 1.2 for new large facilities, after an average of 1.25 in 2025, and Singapore pushes operators to run their halls at 26 °C.

Other levers depend on location. Google Cloud publishes the carbon-free share of electricity for each of its regions (vendor data). Liquid cooling delivers waste heat at 50 to 60 °C, easier to reuse than the 15 to 35 °C from air cooling; a study by KIT finds pipes to a heating network feasible over about 5 km for a large source and 2 to 3 km for small ones. The Uptime Institute expects AI load to push developers towards their own power generation on site, often with gas. Efficiency per task improves while total demand keeps growing, a rebound effect. The energy layer of this stack covers power, grids and water in detail.

Who controls the cloud

The market is concentrated. Synergy Research puts cloud infrastructure revenue at US$143.4 bn in the second quarter of 2026, up 43%, with Amazon at 28%, Microsoft at 20% and Google at 15%. Together the three held 63%; in public IaaS and PaaS, the bulk of the market, their share was 67%. Omdia, which defines the market differently, gave AWS 32% for the fourth quarter of 2025. Amazon, Microsoft and Google also hold 57% of all hyperscale data centre capacity (Synergy, via TechInsights).

Vendors present the build-out as the foundation of a new industry. NVIDIA's Jensen Huang describes AI as a stack of five layers from energy to applications, calls the infrastructure layer AI factories and says the build-out is still at an early stage. Microsoft's Satya Nadella describes a flexible fleet that can switch between training and inference, built from own sites, leases and capacity rented from neoclouds, and avoids gigawatts tied to a single chip generation or a single customer. Both run companies that earn more the more is built.

Critics describe a concentration of power. The economist Cecilia Rikap (UCL) calls Amazon, Microsoft and Google cloud hegemons on which even large corporations and states depend. Jai Vipra and Sarah Myers West of the AI Now Institute describe compute as controlled at a few chokepoints and argue that discounted cloud credits and equity deals tie AI labs to one ecosystem; they propose separating cloud services from chip design. The think tank Open Future puts the combined share of European providers in their home market at about 13%.

Open models change only part of this picture. An open-weight model reduces dependence on one model provider, and it still runs in someone's data centre, as Meredith Whittaker and co-authors point out in Nature.

What this means for your decisions

The following checklist is an Aulory synthesis of the sources on this page. It is written for young companies that use AI and build no data centres of their own.

1. 01Start with APIs. Build prototypes on per-token APIs and measure time to first token, time per output token and cost per task with your real prompts.
2. 02Settle the data question first. Decide which data may leave which jurisdiction. Then choose region and deployment type (regional, data zone or global) and record both in your data processing agreement.
3. 03Commit only to steady load. Reserve capacity or throughput when your usage is steady and predictable. Send work that can wait to batch tiers or spot capacity.
4. 04Keep the exit open. Use open formats, a layer that lets you swap the model provider and exit clauses that match the Data Act deadline of 12 January 2027.
5. 05Self-host for clear reasons. Self-hosting or local AI fits privacy, offline use, high volumes of small models or the need for control. Size power and cooling honestly; GPU racks need far more than an office server room provides.
6. 06Ask for facts. Ask providers for the PUE and WUE of the site, the carbon-free share of the region, heat reuse, the Tier level and credits for outages, and treat answers without a site and a year as marketing.

What the layer is made of

Key figures, with source and date

• 103 to 161 GWof data centre capacity worldwide in 2026, depending on what is counted: JLL about 103 GW of IT capacity, TrendForce 161 GW of power demand.TrendForce; JLL via IREI · 16 Sep 2026 · market research
• 14.2 GWof IT power in the 95 largest tracked AI data centres, with computing power equal to 14.8 million NVIDIA H100s.Epoch AI, Frontier Data Centers · 20 Oct 2026 · neutral
• $143.4 bncloud infrastructure revenue in Q2 2026, up 43%. Amazon, Microsoft and Google together held 63%, and 67% in public IaaS and PaaS.Synergy Research · 30 Jul 2026 · market research
• $115.9 bninvested by Microsoft in property and equipment in its fiscal year to June 2026. Amazon reported $173.0 bn gross over twelve months.Microsoft; Amazon earnings releases · 29 Jul 2026 · self-reported
• 11 vs 142 kWper rack: the most frequently reported density in data centres in 2026, against up to 142 kW for a current NVIDIA rack with 72 GPUs.Uptime Institute and NVIDIA via Data Center Knowledge · 28 Aug 2026 · trade press
• about 1%vacancy in North American data centres for the third year running, with more than 66 GW under construction.JLL · 11 Aug 2026 · consultancy
• 3.7 vs 1.5 GWcontracted against active power at the neocloud CoreWeave at the end of June 2026. Most contracted capacity is still being built.CoreWeave via MarkTechPost · 23 Aug 2026 · self-reported
• 48% to 67%share of global data centre capacity held by hyperscale operators, end of 2025 and forecast for 2031.Synergy Research · 7 Apr 2026 · market research
• 466interruptions in 54 days of training Llama 3 405B on up to 16,384 GPUs, 419 of them unexpected. Effective training time stayed above 90%.Meta, The Llama 3 Herd of Models · Jul 2024 · vendor

Regions compared

Data centres follow power, land, capital and customers. The US leads by a wide margin. Outside the US and China, large new hubs are emerging in Johor (Malaysia), India and the Gulf states, and capacity in Latin America grew by 20 to 41% a year, depending on the source. China reports its build-out as computing power in EFLOPS. Brokers count capacity differently, so figures from different sources are not directly comparable. For Africa, InstaDeep's CEO Karim Beguir estimated the continent's share of world computing capacity at about 0.1% (Rest of World, 2025).

Sources: Synergy Research via TechInsights (20 Aug 2026), CBRE (8 Sep 2026; Q1 2026 via w.media, 15 and 17 Jul 2026), JLL (11 Aug 2026), Epoch AI (17 Apr and 20 Oct 2026), MIIT via Unite.AI (8 Sep 2026), CBRE via BeBeez (18 Aug 2026), EuroHPC (30 Jul 2026), DSIT via CMS (6 Feb 2026), UK policy paper via Baker McKenzie (9 Jan 2026), Cushman & Wakefield (2 Jun and 5 Aug 2026), MeitY via TelecomTalk (Mar 2026), CBRE via Real Estate Asia (29 Sep 2026), EDB Singapore (30 May 2024), DatacenterDynamics (28 May 2025, 21 Jan and 12 Mar 2026), JLL via InvestChile (20 Apr 2026), CIPER Chile (2 Feb 2026), Fortren & Company via Nairametrics (27 Jul 2026), Rest of World (27 Mar 2025). Several figures come from secondary reports of the original.

Examples

• The largest AI data centres according to Epoch: Colossus 2 in Memphis, Anthropic and Amazon in Indiana, Microsoft Fairwater in Atlanta, Meta Prometheus in Ohio, OpenAI Stargate in Abilene.
• Since January 2026, AWS runs a European Sovereign Cloud in Brandenburg. The sovereignty is the provider's own claim.
• Telekom and NVIDIA run an Industrial AI Cloud in Munich since February 2026.
• The EU is looking for up to seven sites for AI gigafactories; applications close on 12 November 2026.

Why it matters

This layer decides three things you feel directly: what AI costs you, where your data is processed and how dependent you become. Capacity is scarce, vacancy in North America sits near 1%, and the largest providers spend over US$100 bn a year each, which keeps the market in few hands. Your choice between an API, a hyperscaler, a neocloud or your own hardware sets your price per task, your exposure to foreign law and the cost of leaving.

It also explains where responsibility lies. Every data centre draws on power, water and land in a specific place, from Memphis to Johor to Santiago de Chile. Customers, regulators and employees ask more and more which sites their AI runs in and under which conditions.

Common misconceptions

• An EU region keeps my data out of reach of US law.

  The US Department of Justice says providers under US jurisdiction must answer valid legal process wherever data is stored. EU law adds its own conditions (DOJ CLOUD Act FAQ; EU Data Act).

• Data stored in the EU is also processed in the EU.

  In global AI deployments, prompts and answers may be processed wherever the model runs (Microsoft documentation).

• If there is free space in the data centre, we can host AI there.

  Power and cooling per rack set the limit: 11 kW is typical, a current AI rack draws up to 142 kW (Uptime Institute, NVIDIA via Data Center Knowledge).

• Serverless means GPUs on demand.

  The term means different things per vendor. Amazon's SageMaker Serverless Inference supports no GPUs at all (AWS documentation).

• Tier IV guarantees a fixed uptime percentage.

  The Tier Standard defines redundancy and maintainability by design. The official page names no percentages (Uptime Institute).

• Leaving a hyperscaler costs a fortune in egress fees.

  The EU bans switching charges from 12 January 2027, and AWS already waives exit egress on request. The real switching costs sit in proprietary services and rebuilding (European Commission; AWS).

• Supercomputer rankings show who has the most AI capacity.

  The TOP500 measures double-precision performance for science. AI runs at much lower precision, and the largest commercial AI clusters are missing from the list (TOP500; Epoch AI).

Perspectives from leading experts

Ten perspectives from vendors, analysts, policy research and critical scholarship, from the US, Europe, the Gulf and Latin America. Each summary paraphrases a published work, linked below it. Vendors are marked as such. We found no published work by an African author on this layer that we could open.

• Jensen Huang

  Founder and CEO of NVIDIA, born in Taiwan, USA · vendor

  AI is essential infrastructure built in five dependent layers: energy, chips, infrastructure, models and applications. He calls the infrastructure layer of land, power, cooling and networking AI factories that produce intelligence, names energy the binding constraint and sees the build-out at an early stage.

  AI Is a 5-Layer Cake (NVIDIA Blog) · 10 Mar 2026

• Satya Nadella

  CEO of Microsoft, born in India, USA · vendor

  Azure should run a fungible fleet that serves training, data generation and inference, without gigawatts built for one chip generation or one large lab. Microsoft mixes own builds, leases and capacity rented from neoclouds, and data residency rules such as the EU Data Boundary force it to build regionally and offer sovereign clouds in France and Germany.

  Interview on the Dwarkesh Podcast, with Dylan Patel · 12 Nov 2025

• Tareq Amin

  CEO of Humain, the AI company of Saudi Arabia's Public Investment Fund · vendor, state-owned

  Humain is looking for a US technology group as an equity partner for its data centres and targets 1.9 GW by 2030 and 6.6 GW by 2034, aiming to handle 7% of the world's training and inference by 2030. According to the report, the state will subsidise electricity, and a first 50 MW site with 18,000 NVIDIA GB300 chips is to grow to 500 MW.

  Statements to the Financial Times, reported by DatacenterDynamics · 28 May 2025

• Epoch AI

  Nonprofit research institute on AI trends, USA · independent analysis

  AI infrastructure has been largely opaque, so the team estimates power, land and compute of the largest sites from satellite images, permits and disclosures, in an open database. At launch it found several US sites heading for 1 GW in 2026, build times of one to three and a half years and about US$30 bn per GW of facility power.

  Introducing the Frontier Data Centers Hub · 4 Nov 2025

• Andy Lawrence

  Executive Director of Research, Uptime Institute, UK and USA · industry body

  Compute for large models will concentrate among a few big organisations, and AI load will strain grids and push developers towards on-site generation, often with gas. Digital infrastructure keeps expanding strongly, while uncertainty about how AI will reshape demand makes capacity planning and resilience harder.

  Five Data Center Predictions for 2026 (Uptime Institute press release) · 13 Jan 2026

• Cecilia Rikap

  Associate Professor of Economics, UCL Institute for Innovation and Public Purpose, London · critical research

  Amazon, Microsoft and Google are cloud hegemons with intellectual monopolies over knowledge and data, on which even firms such as Disney, Nike or Walmart depend to use AI, while states lose core functions. Based on 112 interviews and data on venture capital, acquisitions, patents and open source, she argues for technology developed through democratic planning.

  The Rulers: Corporate Power in the Age of AI and the Cloud (Verso) · 4 Aug 2026

• Francesca Bria and co-authors

  Innovation economist and co-founder of the EuroStack initiative, Italy, with Paul Timmers, Andrea Renda and others · policy research

  More than 80% of Europe's digital infrastructure and technologies are imported, which creates systemic vulnerabilities and limits innovation. The report proposes a strategy to build a European digital stack for strategic autonomy, from infrastructure to services.

  EuroStack: A European alternative for digital sovereignty (CEPS) · 13 Feb 2025

• Jai Vipra and Sarah Myers West

  Researchers at the AI Now Institute at the time of publication, India and USA · critical research

  Compute is a core input to large-scale AI alongside data and labour, and its control sits at a few chokepoints: NVIDIA, TSMC and ASML in hardware and AWS, Google Cloud and Azure in the cloud. Discounted compute and equity-linked deals can lock AI labs into one ecosystem, so they recommend antitrust action such as separating cloud services from chip design.

  Computational Power and AI (AI Now Institute) · 27 Sep 2023

• Zuzanna Warso and Aditya Singh

  Researchers at the think tank Open Future, Europe · policy research

  Europe's reliance on AWS, Azure and Google Cloud comes from market choices, policy gaps and inertia, and European providers hold about 13% combined. Cloud is as essential as energy or telecoms, so they call for transparent pricing, structural separation of infrastructure and services and public procurement that creates demand for European providers on a level playing field.

  Cloud infrastructure as a foundation for digital sovereignty (Open Future policy brief) · 19 Mar 2026

• DataCenterBoom! (Instituto Latinoamericano de Terraformación, founded by Paz Peña)

  Research initiative on the social and environmental impact of data centres in Latin America, Santiago de Chile · critical research

  A development bank report pitches Latin America's cheap clean energy and land to data centre investors and treats renewables as a selling point. The post argues that it ignores land, water and climate justice and describes data centres as extractive enclaves that use shared resources while creating few local jobs or supply chains.

  Los centros de datos quieren ser «verdes» y miran a América Latina. ¿Una buena noticia? · 13 Aug 2026

Terms for this layer

• Data CentreA building full of servers, storage and network equipment, with the power and cooling that AI runs on.
• HyperscalerOne of the few tech giants that run data centres at enormous scale and rent out computing power, storage and AI services.
• Cloud ComputingRenting computing power, storage and software over the internet, on demand, instead of running your own machines.
• ComputeThe computing power, and the chips and software behind it, that AI needs to be trained and to run.
• NeocloudA cloud provider that specialises in renting out GPUs for AI, such as CoreWeave, Nebius, Lambda or Crusoe.
• ColocationRenting space, power, cooling and network connections in someone else's data centre, while you bring and run your own servers.
• Edge ComputingComputing in small sites or devices close to where data is created and used, such as factories, shops, vehicles or phones.
• Sovereign CloudA cloud designed so that location, operation and applicable law stay within one jurisdiction, such as the EU.
• IaaS, PaaS and SaaSThe three standard ways to rent cloud computing: raw infrastructure, a platform for your own code, or finished software.
• GPUA graphics processor: a chip that runs thousands of simple calculations at once, which made it the workhorse of AI.
• CPUThe central processor of every computer: a few strong, flexible cores that work through tasks one step after another.
• HBMVery fast memory stacked right next to an AI chip, so its processors never wait for data. In 2025, it was a key bottleneck.
• InterconnectThe fast network that links thousands of GPUs so that they can work on one AI job together.
• Liquid CoolingCooling chips with a liquid that flows through cold plates on the chip or around the whole server, needed for dense AI racks.
• PUEThe standard efficiency figure for data centres: total electricity of the site divided by the electricity that reaches the IT equipment.
• WUEHow much water a data centre uses on site, measured in litres per kWh of IT electricity.
• TrainingThe one-time, very expensive process that sets a model's parameters by showing it vast amounts of data.
• InferenceRunning a finished model to get an answer. What you pay for every time you send a message.
• Time to First TokenHow long an AI model takes to show the first part of its answer. Together with time per output token, it describes how fast a model feels.
• LatencyThe waiting time between sending a request and getting the answer. In AI, speed is a feature.
• TokenThe small chunk of text a model reads and writes. Price, speed and memory limits are all counted in tokens.
• APIThe programmable connection to a model provider. You pay per token and keep more control than in a chat app.
• Spot InstanceSpare cloud capacity sold at a steep discount, which the provider can take back at short notice.
• Cost per TaskWhat one finished piece of work really costs with AI: tokens, tools, retries and the human time to check it.
• Data ResidencyWhere your data is physically stored and processed. It is one part of data sovereignty.
• Data SovereigntyKeeping control over where your data lives, who can access it and under which laws.
• GDPRThe EU's data protection law. It applies whenever AI tools process personal data about people in the EU.
• Egress FeesCharges a cloud provider bills when data leaves its network, for example when you move to another provider.
• Vendor Lock-inWhen switching to another AI provider becomes so costly or painful that you effectively can't.
• Fine-TuningTraining an existing model further on your own examples to change its behaviour. Rarely the first step.
• Data Processing AgreementThe contract that sets out how a provider may handle personal data on your behalf. Required under the GDPR.
• Model ProviderThe company or machine that runs the model for you: a cloud service or your own computer.
• Open-Weight ModelA model whose trained parameters are published, so anyone can download and run it.
• Self-HostingRunning AI models or tools on infrastructure you control, instead of using someone else's service.
• Local AIRunning a model on your own laptop or server. Nothing leaves the machine, but smaller models fit.
• Rebound EffectWhen efficiency makes something cheaper to use, people use more of it, so total consumption can rise.

Sources

1. Energy and AI: Energy demand from AI, International Energy Agency, 2025
2. Data centre electricity use surged in 2025, International Energy Agency, 16 Apr 2026
3. Tier Classification System, Uptime Institute (sells certification), 2026
4. The growing PUE advantage of larger data centers, Uptime Institute, 6 Aug 2026
5. ISO/IEC 30134-2, Power usage effectiveness, ISO/IEC, 2016, replaced 2026
6. AI rack density's real limits: power, cooling, failure risk, Data Center Knowledge, 28 Aug 2026
7. GB200 NVL72, NVIDIA (vendor), 2026
8. Building Meta's GenAI Infrastructure, Meta Engineering (vendor), 12 Mar 2024
9. The Llama 3 Herd of Models, Meta (vendor report), Jul 2024
10. DistServe: Disaggregating Prefill and Decoding for Goodput-optimized LLM Serving, Zhong et al., USENIX OSDI, Jul 2024
11. Best practices for Compute Engine region selection, Google Cloud (vendor), 22 Oct 2026
12. The NIST Definition of Cloud Computing, SP 800-145, Mell, Grance, NIST, 28 Sep 2011
13. Fog Computing Conceptual Model, SP 500-325, NIST, Mar 2018
14. TOP500 list, June 2026, TOP500, Jun 2026
15. Amazon Bedrock pricing, AWS (vendor), 25 Oct 2026
16. Deployment options in Microsoft Foundry, Microsoft Learn (vendor), 2026
17. Deployment types for Foundry Models, Microsoft Learn (vendor), 2026
18. Serverless Inference, AWS SageMaker documentation (vendor), 2026
19. Amazon EC2 Spot Instances, AWS (vendor), 2026
20. Best GPU neoclouds 2026, MarkTechPost, compiling company releases and list prices, 23 Aug 2026
21. Data center GPU pricing 2026, IntuitionLabs (advisory), 2026
22. Global data center power demand capacity, TrendForce, 16 Sep 2026
23. Global data center sector to nearly double to 200 GW (JLL outlook), IREI, 7 Jan 2026
24. AI data center power, Epoch AI, 16 Jan 2026
25. Frontier Data Centers, Epoch AI, 20 Oct 2026
26. Introducing the Frontier Data Centers Hub, Epoch AI, 4 Nov 2025
27. OpenAI Stargate: where the US sites stand, Epoch AI, 17 Apr 2026
28. Hyperscale operators to account for 67% of all data center capacity by 2031, Synergy Research Group, 7 Apr 2026
29. Hyperscalers' global data center capacity, Q1 2026, TechInsights with Synergy, 20 Aug 2026
30. Q2 Cloud Market Passes $143 Billion; Highest Growth Rate in Eight Years, Synergy Research Group, 30 Jul 2026
31. Global cloud infrastructure spending rose 29% in Q4 2025, Omdia, 26 Mar 2026
32. Amazon.com announces second quarter results, Amazon (company release), 30 Jul 2026
33. FY26 Q4 press release, Microsoft (company release), 29 Jul 2026
34. Meta reports second quarter 2026 results, Meta (company release), 29 Jul 2026
35. Alphabet Q2 2026: capex up to $205bn, cloud +82%, The Next Web (press report), 23 Jul 2026
36. Oracle Q1 strong, and so are capital expenditures, Constellation Research (press report), 10 Sep 2026
37. Oracle forecasts $95 bn in capex for FY27, Outlook Business citing Reuters (press report), 11 Jun 2026
38. CoreWeave Q2 earnings call highlights, MarketBeat (company-reported), 11 Aug 2026
39. North America data center report, midyear 2026, JLL (broker), 11 Aug 2026
40. North American data center demand continues to outpace supply, CBRE (broker), 8 Sep 2026
41. FLAP-D markets plagued by grid constraints and moratoriums, w.media citing CBRE, 17 Jul 2026
42. APAC data centre update H1 2026, Cushman & Wakefield (broker), 5 Aug 2026
43. MIIT plan targets 9,800 EFLOPS of intelligent compute by 2030, Unite.AI citing MIIT, 8 Sep 2026
44. AI Gigafactories call, EuroHPC Joint Undertaking, 30 Jul 2026
45. What you need to know about Africa's first AI factory, Damilare Dosunmu, Rest of World, 27 Mar 2025
46. CLOUD Act FAQ, US Department of Justice, 2026
47. CLOUD Act resources, US Department of Justice, 2026
48. Judgment in Case C-311/18 (Schrems II), press release 91/20, Court of Justice of the European Union, 16 Jul 2020
49. EU-US data transfers, European Commission, 2026
50. Data Act explained, European Commission, 2026
51. Cloud providers lose switching fees in January 2027, ACM guidance says, PPC Land (secondary), Oct 2026
52. Free data transfer out to internet when moving out of AWS, AWS News Blog (vendor), 2024-03-05, updated 2026-10-08
53. Opening the AWS European Sovereign Cloud, AWS News Blog (vendor), 14 Jan 2026
54. The new draft European regulation includes a four-level classification system, InCyber, Jun 2026
55. European Commission advances data center sustainability ratings, Beveridge & Diamond, 30 Sep 2026
56. EU Commission adopts data centre sustainability rating scheme, Linklaters, 14 Oct 2026
57. Waste heat from data centres for district heating, Barth et al., KIT, Geothermal Energy, 22 Apr 2026
58. AI Is a 5-Layer Cake, Jensen Huang, NVIDIA (vendor), 10 Mar 2026
59. Satya Nadella: How Microsoft is preparing for AGI, Dwarkesh Podcast, 12 Nov 2025
60. Humain targets 6.6 GW by 2034 with subsidised electricity, Sebastian Moss, DatacenterDynamics, 28 May 2025
61. Five Data Center Predictions for 2026, Uptime Institute via Business Wire, 13 Jan 2026
62. The Rulers: Corporate Power in the Age of AI and the Cloud, Cecilia Rikap, Verso, 4 Aug 2026
63. EuroStack: A European alternative for digital sovereignty, Bria, Timmers, Gernone, Renda, Fischer, Grabova, CEPS, 13 Feb 2025
64. Computational Power and AI, Jai Vipra, Sarah Myers West, AI Now Institute, 27 Sep 2023
65. Cloud infrastructure as a foundation for digital sovereignty, Zuzanna Warso, Aditya Singh, Open Future, 19 Mar 2026
66. Why 'open' AI systems are actually closed, and why this matters, Widder, Whittaker, West, Nature, Nov 2024
67. Los centros de datos quieren ser «verdes» y miran a América Latina, DataCenterBoom!, Instituto Latinoamericano de Terraformación, 13 Aug 2026